AI agent observability
Inspect every step of a run—retrieval, inference, and tool calls—then replay it with a different model or instruction set.
When AI makes a decision, you should be able to see why.
Who observability is for
Anyone who has to explain an agent decision after the fact: an operator debugging a bad ticket, an engineer comparing two models, a reviewer checking that the answer used the approved source. Dashboards that only show latency averages do not answer those questions.
How a trace is used
Open a run. Read the steps in order. If the answer is wrong, you can see whether retrieval, the model, or a tool caused it. Replay the same input with a change you want to test. Evaluation can then score groundedness, instruction adherence, and tool reliability on that evidence.
01REQUEST received
The input that started the run, stored with the execution rather than only in an application log.
02SOURCES retrieved
Which knowledge was fetched, so a fluent answer can be checked against the corpus.
03CONTEXT assembled
The context the model actually saw, including what was left out.
04MODEL inference
Which endpoint ran, with latency and token use on that step.
05TOOL / MCP called
Arguments and results for each granted tool, in order.
06RESULT returned
The payload that came back into the agent before the final response.
07RESPONSE generated
The output, tied to the steps above so you can see why it was produced.
Replay
Don't just watch an execution. Run it again—change the model, instructions, or context—and compare the original and the replay. Replay is supported as a re-run of a past trace with alternate parameters, not as a guarantee that two providers will score the same.
Questions
What is AI agent observability?
It is the ability to inspect every discrete step an agent takes—from input and retrieval through model inference, tool execution, and the final output—with latency, token cost, and source attribution.
What is execution tracing?
An execution trace is the ordered record of those steps for one run. Obliq stores step type, duration, and payloads so you do not reconstruct the run from separate provider logs.
Can I replay an agent execution?
Yes. Replay re-runs a past trace with alternate parameters, such as a different model or instruction set, so you can compare the original and the replay side by side.